AI Glossary

A

Agent

An AI system that can plan, use tools, and take multi-step actions on its own to reach a goal, instead of just answering one question at a time.

AGI (Artificial General Intelligence)

A hypothetical AI that could match or exceed humans across most intellectual tasks, not just one specialty. Today's systems are not AGI.

AI Safety

The research and engineering practice of reducing harm from AI systems, from filtering bad outputs to preventing misuse.

Alignment

Making an AI system reliably do what people actually intend, rather than what a poorly worded objective literally says.

Attention

The mechanism inside transformers that lets a model weigh which parts of the input matter most when producing each new token.

Autoregressive model

A model that generates text one token at a time, predicting each new token from everything it has produced so far.

B

Batch processing

Sending many AI requests together to be processed asynchronously instead of one by one. Providers often discount batch jobs by around 50%.

Benchmark

A standardized test set used to measure and compare model performance, such as coding or math problem suites.

Bias

Systematic unfairness in a model's outputs, usually absorbed from patterns in its training data.

C

Chain-of-thought

A prompting technique where the model works through a problem step by step before giving its final answer, which improves accuracy on reasoning tasks.

Chatbot

A conversational interface to an AI model or a scripted system. Modern chatbots are usually powered by large language models.

Chunking

Splitting long documents into smaller pieces before embedding them, so a retrieval system can find and return just the relevant parts.

Classifier

A model that assigns an input to a category, such as spam vs. not spam or positive vs. negative sentiment.

Completion

The text a language model generates in response to a prompt.

Context window

The maximum amount of text, measured in tokens, that a model can consider at once, including your prompt, the conversation history, and its reply.

Cosine similarity

A math measure of how similar two embeddings are, widely used to rank search results by meaning.

D

Dataset

A collection of examples, text, images, or labeled records, used to train or evaluate a model.

Deep learning

Machine learning built on neural networks with many layers, the foundation of modern AI systems.

Diffusion model

A type of model that generates images by starting from noise and refining it step by step, used by most AI image generators.

Distillation

Training a smaller, cheaper model to imitate a larger one, keeping most of the quality at a fraction of the cost.

E

Embedding

A list of numbers that represents the meaning of text, images, or other data, so that similar things end up numerically close together.

Emergent abilities

Skills that appear in large models without being explicitly trained, often showing up only after models pass a certain size.

Epoch

One full pass through the training dataset during model training.

Evals (evaluations)

Automated tests that score a model or AI feature against expected outcomes, used to catch regressions before shipping.

F

Few-shot learning

Showing the model a handful of worked examples inside the prompt so it copies the pattern, no retraining required.

Fine-tuning

Further training a pre-trained model on your own examples so it gets better at a specific task, style, or domain.

Foundation model

A large model trained on broad data that serves as a general-purpose base for many downstream tasks.

Function calling

A model capability where, instead of plain text, the model returns a structured request for your code to run, such as look up an order or send an email.

G

Generative AI

AI that creates new content, text, images, audio, code, or video, rather than only classifying or ranking existing data.

GPU

A graphics processing unit, the chip type that does the massive parallel math needed to train and run neural networks.

Gradient descent

The core training algorithm: repeatedly nudge the model's weights in the direction that reduces its error.

Grounding

Tying a model's answers to verifiable sources or live data, for example via retrieval or citations, to reduce made-up answers.

Guardrails

Rules, filters, and checks wrapped around a model to block unsafe inputs and outputs, such as PII filters or topic restrictions.

H

Hallucination

When a model states something false with full confidence, such as inventing a citation, a statistic, or an API that does not exist.

Hyperparameter

A setting chosen before training or inference, such as learning rate or temperature, that controls how the model learns or generates.

I

In-context learning

A model's ability to pick up a new task purely from instructions and examples in the prompt, without any weight updates.

Inference

Running a trained model to get outputs, as opposed to training it. Inference cost is what you pay per API call.

Instruction tuning

Fine-tuning a model on instruction-and-answer pairs so it follows directions instead of just continuing text.

J

Jailbreak

A prompt crafted to trick a model into ignoring its safety rules and producing content it should refuse.

K

Knowledge cutoff

The date after which a model has no built-in knowledge, because its training data ends there. Newer events require search or retrieval.

KV cache

A memory optimization that stores attention computations for earlier tokens so the model does not redo them for every new token, making generation much faster.

L

Latency

How long you wait for a model's response, often measured as time to first token and time to full completion.

LLM (Large Language Model)

A neural network with billions of parameters trained on huge amounts of text to understand and generate language. Claude, GPT, and Gemini are LLMs.

LoRA (Low-Rank Adaptation)

A cheap fine-tuning method that trains small add-on matrices instead of the whole model, cutting cost and memory dramatically.

Loss function

The formula that scores how wrong a model's predictions are during training. Training is the process of driving this score down.

M

MCP (Model Context Protocol)

An open standard for connecting AI assistants to external tools and data sources through a common interface, so any MCP client can use any MCP server.

Mixture of Experts (MoE)

A model design where only a few specialized sub-networks activate per token, giving big-model quality at lower compute cost.

Model weights

The billions of learned numbers inside a neural network that encode everything it knows. Training adjusts them; inference just reads them.

Multimodal

A model that handles more than one type of input or output, such as text plus images, audio, or video.

N

Neural network

A computing system of layered, connected nodes loosely inspired by the brain, which learns patterns from data.

NLP (Natural Language Processing)

The field of AI focused on understanding and generating human language.

O

Open weights

A model whose trained weights are published so anyone can download, run, and fine-tune it, such as Llama or Mistral models.

Orchestration

Coordinating multiple AI calls, tools, and business logic into one workflow, for example a pipeline that retrieves, summarizes, then drafts an email.

Overfitting

When a model memorizes its training data instead of learning general patterns, so it performs well in training but poorly on new inputs.

P

Parameters

The individual learned values inside a model. "A 70B model" means 70 billion parameters; more usually means more capable but more expensive.

Perplexity

A measure of how surprised a model is by text; lower perplexity means the model predicts the text better.

Pre-training

The initial, massive training phase where a model learns language and world knowledge from broad data, before any task-specific tuning.

Prompt

The input you give a model: the question, instructions, and any context or examples.

Prompt caching

An API feature that stores the processed form of a repeated prompt prefix so later requests reuse it, cutting cost by up to 90% and reducing latency.

Prompt engineering

The practice of writing and structuring prompts to get consistently better results, including roles, examples, and output format instructions.

Prompt injection

An attack where malicious instructions hidden in content the model reads, like a web page or email, hijack the model's behavior.

Q

Quantization

Shrinking a model by storing its weights in lower precision numbers, making it smaller and faster with a small quality trade-off.

Query

The question or search text sent to a model or retrieval system.

R

RAG (Retrieval-Augmented Generation)

A pattern where the system first retrieves relevant documents, then has the model answer using them, so responses reflect your data instead of only training memory.

Reasoning model

A model that spends extra compute "thinking" through a problem internally before answering, improving results on hard math, code, and planning tasks.

Red teaming

Deliberately attacking your own AI system, with jailbreaks, injections, and edge cases, to find failures before real users or attackers do.

Reinforcement learning

Training by trial and error: the model takes actions, gets rewards or penalties, and learns to prefer actions that score well.

RLHF (Reinforcement Learning from Human Feedback)

Training a model against human preference ratings so its answers become more helpful, honest, and safe. A key step in building assistant models.

S

Semantic search

Search by meaning instead of exact keywords, powered by embeddings, so "reset my password" matches "credential recovery".

Streaming

Receiving a model's answer token by token as it generates, instead of waiting for the whole response. This is how chat UIs show text appearing live.

Structured outputs

Forcing a model to reply in a guaranteed format, usually JSON matching a schema, so software can reliably parse the result.

Supervised learning

Training on labeled examples, input plus correct answer, so the model learns to map one to the other.

Synthetic data

Training or test data generated by AI rather than collected from the real world, used to fill gaps or protect privacy.

System prompt

Hidden instructions given to a model before the user's message that set its role, rules, and style, such as "You are a helpful support agent for Acme".

T

Temperature

A setting that controls randomness in generation: low values give focused, repeatable answers; high values give more varied, creative ones.

Token

The unit of text a model reads and writes, roughly 3-4 characters or about three-quarters of an English word. API pricing is per token.

Tokenizer

The component that splits text into tokens before the model processes it. Different models use different tokenizers, so counts vary.

Tool use

A model's ability to call external tools, search, code execution, databases, APIs, during a conversation to get things done or fetch facts.

Top-p (nucleus sampling)

A sampling setting that limits generation to the smallest set of likely next tokens whose combined probability reaches p, trimming off unlikely words.

Training data

The text, images, or examples a model learns from. Its quality and coverage largely determine what the model can do.

Transfer learning

Reusing what a model learned on one task as the starting point for another, the idea behind pre-train then fine-tune.

Transformer

The neural network architecture, built on attention, that powers virtually all modern language models. Introduced in the 2017 paper "Attention Is All You Need".

U

Unsupervised learning

Learning patterns from data without labels, such as clustering similar documents or predicting the next word in raw text.

V

Vector

An ordered list of numbers. In AI, embeddings are vectors that place meaning in a mathematical space.

Vector database

A database built to store embeddings and find the most similar ones quickly, the storage layer behind most RAG and semantic search systems.

Z

Zero-shot learning

Asking a model to do a task with instructions only and no examples, relying entirely on what it learned in training.